Efficient multi-stage alignment of multispectral images using a multi-GPU algorithm

Accurate registration of high-resolution multispectral UAV orthomosaics acquired on different dates is critical for remote sensing applications such as change detection or land cover monitoring. This task is challenging due to variability in acquisition conditions, including seasonal changes, illumination differences, weather variability, and sensor characteristics. This article presents a parallel multilevel registration method that combines HSI-KAZE (Hyperspectral KAZE), a feature-based approach for robust coarse alignment, with Hyperspectral Fourier-Mellin (HYFM), an area-based method for registration refinement. The method consists of three progressive registration levels that produce: coarse scale and rotation correction in the first level, scale, rotation and shift adjustment in the second one, and final shift refinement in the third one. The parallel implementation of the proposed method, using MPI, OpenMP, and CUDA, efficiently processes multiple datasets on GPU-accelerated high-performance computing (HPC) clusters. Experiments on four pairs of multitemporal multispectral orthomosaics from river environments demonstrate high registration accuracy. The final multi-node and multi-GPU implementation shows an speedup of 20.94$\times$ compared to the OpenMP implementation.

keywords: Image registration, Remote Sensing, multispectral imaging, High Performance Computing, Parallel Computing